Skip to course content
Free RAG fundamentals course

Introduction to RAG and Knowledge Assistants

Unit 05.03: Answers that cite the wrong chunk convincingly

The hardest failure to catch is an answer with a real citation, pointing at a real chunk, making a claim that chunk does not support.

Comparing the claim against its evidence

Every superficial check passes here. There is a citation. The id resolves. The chunk exists and is on topic. The only thing wrong is the relationship between the claim and the text it points at.

The example below takes an answer claiming a 30-day refund window, cited to a chunk about 30-day *exchanges*, and tests whether the cited text supports it.

CHUNKS
  c1   Refunds are allowed within 7 days of purchase.
  c2   Exchanges are allowed within 30 days of purchase.

ANSWER      "Refunds are allowed within 30 days."     cites: c2
CITED TEXT  "Exchanges are allowed within 30 days of purchase."

supported by its citation?   NO
terms in the claim that are not in the cited text:   refunds

The citation is real, the chunk exists, and the number came from a document
about something else. Every check that only asks "is there a citation?"
passes this.

The check is deliberately crude: pull the substantive terms out of the claim and ask which of them appear in the cited chunk. The missing terms are printed, and they are the ones carrying the error.

Term overlap is not entailment, and this check will miss subtler mismatches. It also runs in microseconds on every answer, which a sophisticated check will not, and it catches the common case where a number or a noun came from somewhere other than the citation.

The mistake this prevents

The mistake is measuring citation *coverage* - what fraction of claims have a citation - and calling it grounding. Coverage is trivially gamed by attaching an id to everything. What matters is whether the cited chunk contains the claim, which is a different measurement entirely.

Takeaway

A citation is a claim about provenance and needs verifying like any other claim. Compare the answer's substantive terms against the cited text automatically, on every answer, and inspect what gets flagged.